AI Architecture Shifts: TabFM, Inkling, and Power Efficiency

Three distinct architecture bets are reshaping how AI models are built and deployed in July 2026: Google Research published TabFM, a zero-shot foundation model for tabular data; Thinking Machines open-sourced Inkling, a 975-billion-parameter multimodal Mixture-of-Experts model; and NVIDIA’s infrastructure blog made the case that performance per watt — not raw compute — is the metric that determines which AI factories survive a power-constrained scaling race.

Google TabFM Skips Per-Dataset Training for Tables

Google Research’s TabFM treats tabular prediction as an in-context learning problem, generating predictions for a new, unseen dataset in a single forward pass — no retraining, no hyperparameter search. According to Google Research’s blog, the model can reduce time-to-production from weeks of pipeline engineering to a single API call.

Traditional gradient-boosted tree models require data scientists to clean inputs, impute missing values, encode categorical variables, and run repetitive hyperparameter optimization loops across learning rates, tree depths, subsampling ratios, and regularization grids. Once deployed, those models still demand ongoing maintenance. “Traditional models incur ongoing operational debt through data drift monitoring and retraining pipelines to stay accurate,